Real-time markerless motion capture for Blender: a webcam streams full-body SMPL-X poses straight onto your character. Built in collaboration with Corridor Digital.
Free and open source.
Point the engine at a video and it streams live SMPL-X poses. Here it drives the standalone skeleton viewer on two of the repo's test clips — source on the left, the streamed pose on the right:
These clips are the repo's own test fixtures (tests/fixtures/video/), run through the real capture pipeline. The viewer draws approximate bone lengths; inside Blender the same stream drives a full SMPL-X character, shown step by step in the Getting Started guide.
- Download the Windows installer from the latest release and run it — no administrator rights needed. The open-source installer is intentionally unsigned; read the SmartScreen and download-verification steps before running it.
- Follow the Getting Started guide — it takes you from a clean machine to a character moving on your webcam, with a screenshot at every step. About 20 minutes, most of it downloads you can leave running.
You are guided inside Blender too: a Getting Started checklist sits at the top of the PoseCap panel and keeps the capture buttons disabled until you are ready, so you can't click into an error.
| Guide | What it covers |
|---|---|
| Getting started | The full walk-through: install → body models → character → live capture |
| Set up the body models | The one-time SMPL-X download, with your own free research account |
| Set up a character | Bring in a Mixamo or Unreal character and convert it in one click |
| Live capture | Stream from a webcam or a video, and record the motion to keyframes |
Early development — first installable preview. See the PRD for scope and roadmap.
Two components, one pipeline:
- Blender extension — panels and operators inside Blender: start/stop the live stream, record mocap to the timeline, capture single poses, manage SMPL-X body models.
- Engine bridge — a separate GPU process wrapping the PEAR pose-estimation model (single image in, SMPL-X parameters out, real-time rates). Streams poses to Blender over a local socket.
Blender animators who want believable body animation without mocap suits, markers, or external mocap software. Corridor Digital's production artists are the founding users; if you have Blender and a webcam, the MediaPipe Lite backend gives you an account-free CPU path. PEAR remains available for the higher-fidelity NVIDIA workflow.
| Component | Supported now |
|---|---|
| OS | Windows 10 / 11 |
| MediaPipe Lite | CPU-first body capture. No GPU required |
| PEAR | NVIDIA/CUDA, driver R570 or newer; qualified on an RTX 3080. RTX 50-series (Blackwell) supported since v1.0.7; GTX 10-series (Pascal) and older are not |
| Blender | 4.2 LTS minimum, 5.x supported |
| Python | 3.11 (Blender bundled for the extension; uv-managed venv for the engine bridge) |
| Camera | Any webcam, including virtual cameras (e.g. Iriun) |
Other PEAR GPU generations remain unqualified unless a release note says otherwise. Linux support for the engine bridge is on the roadmap; macOS is not currently planned.
Two processes, joined by explicit contracts:
- The engine captures webcam frames, finds the person and estimates their full-body pose on the GPU, and streams those poses to Blender over a local connection.
- The Blender extension consumes the stream on a background thread and applies poses on the main thread as they arrive — without wiping your existing keyframes.
- Poses apply pelvis-locked: monocular depth estimation cannot recover trustworthy world position, so world translation stays out until a solid software approach lands (camera tracking is the leading candidate — see the roadmap).
Step-by-step diagrams for every flow (live streaming, capture jobs, install) live in doc/workflows.md; binding structure lives in ARCHITECTURE.md.
Two separate licenses apply — the plugin's and the body models':
- PoseCap itself is free and open source: GPL-3.0 for the Blender extension (required for Blender API linkage), Apache-2.0 for the contracts, core, and engine-bridge libraries — decided in ADR-0006.
- The SMPL, SMPL-X, and FLAME body models are licensed by the Max Planck Institute for research (non-commercial) use. They are never distributed with PoseCap — not in this repo, not in the installer. Each user registers on the official MPI sites and accepts the license terms personally; the plugin's setup wizard then automates the download using that user's own account credentials. The illustrated setup guide walks through it.
- Commercial production use of the body models requires a commercial license from Meshcapade, independent of the plugin itself being free.
The product roadmap lives in the PRD, the single owner of scope and sequence.
- ARCHITECTURE.md — layers, boundaries, IPC contracts
- doc/workflows.md — sequence diagrams and flowcharts per functionality
- AGENTS.md — operational guide for contributors and coding agents
- doc/product/PRD.md — product scope, metrics, roadmap
- doc/reference/ — the papers and upstream projects this builds on
See CONTRIBUTING.md — setup from clone to green tests, the project map, commit conventions, and the rules CI enforces. Short version: most of the codebase (contracts/, core/) needs no GPU and no Blender to work on; never commit model files or weights; one concern per signed-off commit.
- Dean / Corridor Digital — concept and the production use case driving this
- Alê Alvaro (@alexandremendoncaalvaro) — the PoseCap rewrite: architecture, implementation, and installers
- PEAR (Wu et al., IDEA) — the pose-estimation backbone
- Fast SAM 3D Body (Yang et al.) — roadmap engine backend
- Meshcapade / MPI SMPL Blender addon — the addon lineage the POC started from
- SMPL-X body model by MPI for Intelligent Systems
